strategy: improve harmonic by adding HMM filter to denoise shark signal

strategy: improve harmonic by adding HMM filter to denoise shark signal
This commit is contained in:
austin362667 2022-10-07 00:33:18 +08:00 committed by Austin Liu
parent b80ac89486
commit c8aa4ae400
2 changed files with 118 additions and 49 deletions

View File

@ -16,8 +16,8 @@ exchangeStrategies:
harmonic:
symbol: BTCBUSD
interval: 1s
window: 500
quantity: 0.05
window: 60
quantity: 0.005
# Draw pnl
drawGraph: true
graphPNLPath: "./pnl.png"
@ -26,12 +26,12 @@ exchangeStrategies:
backtest:
sessions:
- binance
startTime: "2022-09-30"
endTime: "2022-10-01"
startTime: "2022-10-01"
endTime: "2022-10-07"
symbols:
- BTCBUSD
accounts:
binance:
balances:
BTC: 1.0
BUSD: 40_000.0
BUSD: 60_000.0

View File

@ -3,14 +3,17 @@ package harmonic
import (
"context"
"fmt"
"os"
"sync"
"github.com/c9s/bbgo/pkg/bbgo"
"github.com/c9s/bbgo/pkg/data/tsv"
"github.com/c9s/bbgo/pkg/datatype/floats"
"github.com/c9s/bbgo/pkg/fixedpoint"
"github.com/c9s/bbgo/pkg/indicator"
"github.com/c9s/bbgo/pkg/types"
"github.com/sirupsen/logrus"
floats2 "gonum.org/v1/gonum/floats"
)
const ID = "harmonic"
@ -317,6 +320,7 @@ func (s *Strategy) Run(ctx context.Context, orderExecutor bbgo.OrderExecutor, se
}
})
s.InitDrawCommands(&profitSlice, &cumProfitSlice)
s.orderExecutor.TradeCollector().OnPositionUpdate(func(position *types.Position) {
bbgo.Sync(ctx, s)
})
@ -332,66 +336,131 @@ func (s *Strategy) Run(ctx context.Context, orderExecutor bbgo.OrderExecutor, se
if klines, ok := kLineStore.KLinesOfInterval(s.shark.Interval); ok {
s.shark.LoadK((*klines)[0:])
}
states := types.NewQueue(s.Window)
states.Update(0)
s.session.MarketDataStream.OnKLineClosed(types.KLineWith(s.Symbol, s.Interval, func(kline types.KLine) {
log.Infof("Shark Score: %f, Current Price: %f", s.shark.Last(), kline.Close.Float64())
//previousRegime := s.shark.Values.Tail(10).Mean()
//zeroThreshold := 5.
nextState := alpha(s.shark.Array(s.Window), states.Array(s.Window), s.Window)
states.Update(nextState)
log.Infof("Denoised signal via HMM: %f", states.Last())
if s.shark.Rank(s.Window).Last()/float64(s.Window) > 0.99 { // && ((previousRegime < zeroThreshold && previousRegime > -zeroThreshold) || s.shark.Index(1) < 0)
if s.Position.IsShort() {
_ = s.orderExecutor.GracefulCancel(ctx)
s.orderExecutor.ClosePosition(ctx, fixedpoint.One, "close short position")
if states.Length() < s.Window {
return
}
_, err := s.orderExecutor.SubmitOrders(ctx, types.SubmitOrder{
Symbol: s.Symbol,
Side: types.SideTypeBuy,
Quantity: s.Quantity,
Type: types.OrderTypeMarket,
Tag: "shark long: buy in",
})
if err == nil {
_, err = s.orderExecutor.SubmitOrders(ctx, types.SubmitOrder{
Symbol: s.Symbol,
Side: types.SideTypeSell,
Quantity: s.Quantity,
Price: fixedpoint.NewFromFloat(s.shark.Highs.Tail(100).Max()),
Type: types.OrderTypeLimit,
Tag: "shark long: sell back",
})
}
if err != nil {
log.Errorln(err)
}
} else if s.shark.Rank(s.Window).Last()/float64(s.Window) < 0.01 { // && ((previousRegime < zeroThreshold && previousRegime > -zeroThreshold) || s.shark.Index(1) > 0)
direction := 0.
if s.Position.IsLong() {
_ = s.orderExecutor.GracefulCancel(ctx)
s.orderExecutor.ClosePosition(ctx, fixedpoint.One, "close long position")
direction = 1.
} else if s.Position.IsShort() {
direction = -1.
}
_, err := s.orderExecutor.SubmitOrders(ctx, types.SubmitOrder{
if s.Position.IsOpened(kline.Close) && states.Mean(5) == 0 {
s.orderExecutor.ClosePosition(ctx, fixedpoint.One)
}
if states.Mean(5) == 1 && direction != 1 {
_, _ = s.orderExecutor.SubmitOrders(ctx, types.SubmitOrder{
Symbol: s.Symbol,
Side: types.SideTypeBuy,
Quantity: s.Quantity,
Type: types.OrderTypeMarket,
Tag: "shark long",
})
} else if states.Mean(5) == -1 && direction != -1 {
_, _ = s.orderExecutor.SubmitOrders(ctx, types.SubmitOrder{
Symbol: s.Symbol,
Side: types.SideTypeSell,
Quantity: s.Quantity,
Type: types.OrderTypeMarket,
Tag: "shark short: sell in",
Tag: "shark short",
})
if err == nil {
_, err = s.orderExecutor.SubmitOrders(ctx, types.SubmitOrder{
Symbol: s.Symbol,
Side: types.SideTypeBuy,
Quantity: s.Quantity,
Price: fixedpoint.NewFromFloat(s.shark.Lows.Tail(100).Min()),
Type: types.OrderTypeLimit,
Tag: "shark short: buy back",
})
}
if err != nil {
log.Errorln(err)
}
}
}))
bbgo.OnShutdown(ctx, func(ctx context.Context, wg *sync.WaitGroup) {
defer wg.Done()
// Output accumulated profit report
if bbgo.IsBackTesting {
defer s.AccumulatedProfitReport.Output(s.Symbol)
if s.DrawGraph {
if err := s.Draw(&profitSlice, &cumProfitSlice); err != nil {
log.WithError(err).Errorf("cannot draw graph")
}
}
}
_, _ = fmt.Fprintln(os.Stderr, s.TradeStats.String())
_ = s.orderExecutor.GracefulCancel(ctx)
})
return nil
}
// TODO: dirichlet distribution is a too naive solution
func observationDistribution(y_t, x_t float64) float64 {
if x_t == 0. && y_t == 0 {
// observed zero value from indicator when in neutral state
return 1.
} else if x_t > 0. && y_t > 0. {
// observed positive value from indicator when in long state
return 1.
} else if x_t < 0. && y_t < 0. {
// observed negative value from indicator when in short state
return 1.
} else {
return 0.
}
}
func transitionProbability(x_t0, x_t1 int) float64 {
// stick to the same sate
if x_t0 == x_t1 {
return 0.99
}
// transit to next new state
return 1 - 0.99
}
func alpha(y_t []float64, x_t []float64, l int) float64 {
al := make([]float64, l)
an := make([]float64, l)
as := make([]float64, l)
long := 0.
neut := 0.
short := 0.
// n is the incremental time steps
for n := 2; n <= len(x_t); n++ {
for j := -1; j <= 1; j++ {
sil := make([]float64, 3)
sin := make([]float64, 3)
sis := make([]float64, 3)
for i := -1; i <= 1; i++ {
sil = append(sil, x_t[n-1-1]*transitionProbability(i, j))
sin = append(sin, x_t[n-1-1]*transitionProbability(i, j))
sis = append(sis, x_t[n-1-1]*transitionProbability(i, j))
}
if j > 0 {
long = floats2.Max(sil) * observationDistribution(y_t[n-1], float64(j))
al = append(al, long)
} else if j == 0 {
neut = floats2.Max(sin) * observationDistribution(y_t[n-1], float64(j))
an = append(an, neut)
} else if j < 0 {
short = floats2.Max(sis) * observationDistribution(y_t[n-1], float64(j))
as = append(as, short)
}
}
}
maximum := floats2.Max([]float64{long, neut, short})
if maximum == long {
return 1
} else if maximum == short {
return -1
}
return 0
}